[nickajacks1]

The existing implementation used a quadratic search for completions, which becomes noticeably slow when an app emits 1000+ completion candidates. Update to a single-pass approach and remove an extra subshell per reply. Regression tested on bash-3.2.

On a real app that emits 1278 candidates, speed is improved from roughly 20 seconds to basically instantaneous.

What type of PR is this?

• cleanup

What this PR does / why we need it:

Bash completion enters the tens of seconds when an app emits 1000+ replies.

The single pass also stops the script from expanding candidates. It used to pass them to compgen -W, which expands each word, so a candidate such as $(touch${IFS}x) ran as a command when the user pressed TAB. The second commit adds a test for this.

Which issue(s) this PR fixes:

I didn't make one :(

Testing

Tested on bash 5 and bash 3.2

Release Notes

Improve performance of bash completion. Bash completion no longer expands candidates, so a candidate such as `$(cmd)` no longer runs as a command.

[Juneezee]

LGTM, thanks!

I reproduced the slowdown with 1278 candidates, the count in the PR description. Listing them (the second TAB, COMP_TYPE=63) took 12.43s on main and 0.12s with this PR. The first TAB was already fast on both, so only the listing was slow.

To reproduce, save this as repro.sh and run bash repro.sh from the root of a urfave/cli checkout, once on main and once on this PR (gh pr checkout 2514):

# A fake app that prints 1278 completion candidates with descriptions.
app() { for ((i = 1; i <= 1278; i++)); do echo "cmd$i:description $i"; done; }

# Stub for bash-completion's _comp_initialize, so this runs as a plain script.
_comp_initialize() { cur=${COMP_WORDS[COMP_CWORD]}; words=("${COMP_WORDS[@]}"); cword=$COMP_CWORD; }

# Render the template the way `app completion bash` does.
eval "$(sed 's/%\[1\]s/app/g; s/%%/%/g' autocomplete/bash_autocomplete)"

# Complete `app <TAB>`: COMP_TYPE 9 is the first TAB, 63 the second (the listing).
COMP_WORDS=(app '') COMP_CWORD=1
for COMP_TYPE in 9 63; do
  echo "COMP_TYPE=$COMP_TYPE"
  time __app_bash_autocomplete
  echo "${#COMPREPLY[@]} replies, first: ${COMPREPLY[0]}"
done

The real times on my machine (bash 5.3):

       COMP_TYPE=9  COMP_TYPE=63
main   0.10s        12.43s
PR     0.09s        0.12s

For ordinary candidates the replies are the same as before: both TABs, with and without descriptions, with and without a typed prefix. A double TAB in an interactive bash 5.3 lists the same thing.

The single pass also brings three small improvements:

• Descriptions are padded to the longest match, not the longest candidate overall.
• A repeated token shows its own description, not the first one's.
• The token and description arrays are now local, not globals left in the user's shell.

cli/autocomplete/bash_autocomplete

Lines 44 to 61 in db1ab19

	local -a matches=() descriptions=()
	local line token description
	local longest=0
	while IFS=$'\n' read -r line; do
	token="${line%%%%:*}"
	if [[ -z "${token}" || "${token}" != "${cur}"* ]]; then
	continue
	fi
	description=""
	if [[ "${line}" == *:* ]]; then
	description="${line#*:}"
	fi
	matches+=("${token}")
	descriptions+=("${description}")
	(( ${#token} > longest )) && longest=${#token}
	done <<< "${opts}"